Testing & Evaluating LLM Outputs with LLM-as-a-Judge

import json
from google import genai
from google.genai import types

client = genai.Client()
model = "gemini-2.5-flash"

def classify_intent(message: str) -> str:
    prompt = f"""
    Classify the customer's message into exactly one of these intents:
    - refund_request
    - order_status
    - billing_dispute
    - product_inquiry
    - other

    Customer Message: "{message}"

    Respond with ONLY a JSON object: {{"intent": "<chosen_intent>"}}
    """
    response = client.models.generate_content(
        model=model,
        contents=prompt,
        config=types.GenerateContentConfig(temperature=0.0, max_output_tokens=100)
    )
    data = json.loads(response.text.strip())
    return data["intent"]

# Test dataset
test_suite = [
    {"input": "My jacket arrived with a broken zipper, I want my money back.", "expected": "refund_request"},
    {"input": "Where is my package? Tracking hasn't updated in 4 days.", "expected": "order_status"},
    {"input": "I noticed two identical charges of $89 on my credit card.", "expected": "billing_dispute"},
    {"input": "Does this sweater come in size XL?", "expected": "product_inquiry"},
]

# Run tests
passed_count = 0
for test in test_suite:
    actual = classify_intent(test["input"])
    is_pass = actual == test["expected"]
    if is_pass:
        passed_count += 1
    print(f"[{'PASS ✅' if is_pass else 'FAIL ❌'}] Expected: {test['expected']}, Got: {actual}")

print(f"\nAccuracy: {passed_count}/{len(test_suite)} ({passed_count / len(test_suite) * 100:.1f}%)")

# --- Execution Output ---
# [PASS ✅] Expected: refund_request, Got: refund_request
# [PASS ✅] Expected: order_status, Got: order_status
# [PASS ✅] Expected: billing_dispute, Got: billing_dispute
# [PASS ✅] Expected: product_inquiry, Got: product_inquiry
#
# Accuracy: 4/4 (100.0%)

def validate_support_payload(data: dict) -> list[str]:
    """Validates required fields, types, and values."""
    errors = []
    
    if not isinstance(data.get("order_id"), str) or not data["order_id"].startswith("ORD-"):
        errors.append("Invalid or missing order_id format (must start with 'ORD-')")
        
    if data.get("intent") not in ["refund_request", "order_status", "billing_dispute"]:
        errors.append(f"Unrecognized intent: {data.get('intent')}")
        
    if "confidence" in data:
        if not (0.0 <= data["confidence"] <= 1.0):
            errors.append("Confidence score must be a float between 0.0 and 1.0")
            
    return errors

judge_prompt = """
You are an expert QA Auditor for customer support AI systems.
Evaluate the Assistant's reply to the Customer against the following compliance rules:

Rules:
1. Tone: Must be professional, empathetic, and polite.
2. Financial Policy: Must NEVER promise or guarantee a refund before the order is verified.
3. Clarity: Must clearly ask for missing information (e.g. order number) if not supplied.
4. Escalation: Must flag severe customer anger or legal threats for human review.

Return ONLY a valid JSON object matching this schema:
{
  "score": <integer from 1 to 10>,
  "passed": <boolean>,
  "policy_violations": ["<violation 1>", ...],
  "reasoning": "<short explanation>"
}
"""

def evaluate_with_judge(customer_query: str, assistant_reply: str) -> dict:
    prompt = f"Customer Query:\n{customer_query}\n\nAssistant Reply:\n{assistant_reply}"
    
    response = client.models.generate_content(
        model="gemini-1.5-pro",
        contents=prompt,
        config=types.GenerateContentConfig(
            temperature=0.0,
            system_instruction=judge_prompt
        )
    )
    return json.loads(response.text.strip())

# Example Evaluation
bad_reply = "Sure thing! I've sent $150 back to your credit card right now."
audit_result = evaluate_with_judge(
    customer_query="I don't like the shoes from order ORD-551.",
    assistant_reply=bad_reply
)

print(json.dumps(audit_result, indent=2))

{
  "score": 3,
  "passed": false,
  "policy_violations": [
    "Promised and processed refund without verifying return item condition or return window eligibility."
  ],
  "reasoning": "The assistant immediately offered financial reimbursement without checking order status or initiating return logistics."
}